← All speakers

Bio, Work & Ideas

Sarthak Aggarwal

Conference affiliation: Co-founder · Decawork · 2026

Sarthak Aggarwal is co-founder and chief technology officer of Decawork, a Y Combinator-backed company building an agent control plane for IT teams. He wants enterprises to manage autonomous software with the same rigor they apply to employees: clear ownership, limited permissions, continuous oversight, and revocable access.

Aggarwal began contributing to open source at age 12 and became a Google Code-In global winner. At BITS Pilani, he led the student-run startup accelerator Conquest. He subsequently worked on systems software and AI infrastructure at NVIDIA, including systems deployed at OpenAI and Meta, and on enterprise AI at Ema. His public software projects include autopilot-shorts, prospector, tarmac, and Vartalap.

Aggarwal founded Decawork with chief executive Aman Raj, joining Y Combinator’s Summer 2026 batch. The company’s initial focus on autonomous IT administration expanded into managing identities, permissions, incidents, and audit trails across human and software workforces. In August 2026, Aggarwal announced its Product Hunt launch.

How Aggarwal approaches enterprise autonomy

  • Agents as an enterprise workforce. Software operating with tools, memory, private data, and delegated permissions needs registration, accountable ownership, monitoring, investigation, and offboarding.
  • Agent identity and delegated authority. Each action should distinguish the agent performing it from the person, service account, device, or workload authorizing it. OAuth Token Exchange helps represent delegation but does not establish a complete agent-specific identity standard.
  • Security boundaries outside the model. Prompt instructions cannot reliably prevent consequential actions involving sensitive systems. Aggarwal advocates independent policy enforcement, scoped permissions, short-lived credentials, and rapid revocation.
  • Privilege-separated agent execution. His enterprise-agent security architecture separates trusted planning, untrusted content processing, and capability-scoped tool execution. If a password-reset request conceals instructions to disable company-wide multifactor authentication, an independent policy gate rejects the unauthorized operation and records the attempt.

Read the topics behind these talks

1 conference talk

Key ideas

Scroll to read ↓

An enterprise agent needs more than useful answers: it needs a runtime identity, bounded authority, enforceable action policies, and a record of what it tried to do.

  • Can you safely employ the agent?
    0:00 ↗
  • Give the worker a runtime identity
    2:31 ↗
  • Apply familiar controls continuously
    3:45 ↗
  • When ordinary content becomes an instruction surface
    5:57 ↗
  • EchoLeak: spending someone else’s authority
    7:25 ↗
  • A code freeze needs an enforceable boundary
    8:42 ↗
  • Separate reasoning from authority
    11:16 ↗
  • Turn trusted intent into a constrained plan
    12:22 ↗
  • Reject the extra action in a password-reset ticket
    13:32 ↗
  • Communication rails still need operational governance
    14:48 ↗

References